source stringclasses 565
values | kind stringclasses 6
values | target stringlengths 10 76 | conditions stringlengths 0 683 |
|---|---|---|---|
belouadah2020-scail | argues-for | initial-state-statistics-rectify-past-class-scores | class-IL from scratch, ILSVRC/VGGFace2/Landmarks/CIFAR-100, Z=10, B≈0.5–2%: classifier weights stored at a class's initial well-trained state are reused and rank-scaled from stored weight statistics with no held-out data or retraining; ScaILherd improves the FT backbone by 2.69 GIL points, the best of all methods teste... |
belouadah2020-scail | argues-for | task-recency-bias-dominates-class-il-errors | class-IL from scratch, ILSVRC Z=10: raw fine-tuning scores are biased toward recent new classes — Fig. 2/3 show past-class mean scores trailing new — and a purely classifier-level rescaling, leaving the representation untouched, rebalances predictions and beats iCaRL/BiC, GIL −6.40 vs iCaRL −16.75 |
belouadah2020-scail | argues-for | old-new-data-imbalance-drives-class-il-degradation | same setup: the new-class bias is attributed to the imbalance a bounded memory forces — past classes re-learned from few exemplars get weaker classifiers than at their initial full-data state; ScaIL's advantage is largest at the smallest memory B, the most imbalanced case |
belouadah2020-scail | nuances | knowledge-distillation-preserves-past-knowledge | class-IL from scratch with exemplar memory, large scale: ablating the distillation term and using vanilla fine-tuning is a stronger backbone — FT_distill makes ~3× more past-past errors e(p,p) than FT at ILSVRC Z=10 B=5000; distillation is found useful only at small scale, hurting large imbalanced datasets with memory |
belouadah2020-scail | builds-on | belouadah2019-il2m | |
belouadah2020-scail | builds-on | rebuffi2017-icarl | |
belouadah2020-scail | builds-on | belouadah2018-deesil | |
lesort2019-regularization-shortcomings | argues-for | regularization-alone-fails-in-class-IL | class-IL single-head, from scratch; impossibility argument — the old-class/new-class inequality is never estimable — plus MNIST-Fellowship 3×10 classes, where EWC/EWC-KFAC/OGD track the fine-tuning baseline while rehearsal with 100 samples/class succeeds |
lesort2019-regularization-shortcomings | argues-against | protecting-important-weights-prevents-forgetting | class-IL single-head; weight protection preserves past intra-task boundaries but cannot optimize inter-task boundaries, so it is "not sufficient to ensure continual learning" — holds for any weight-protection scheme, shown for a 2-task linear minimal example |
lesort2019-regularization-shortcomings | nuances | protecting-important-weights-prevents-forgetting | multi-head/task-IL MNIST-Fellowship: EWC/EWC-KFAC/OGD significantly beat fine-tuning — the weight-protecting role itself works; what regularization cannot replace is the task label's inter-task discrimination |
lesort2019-regularization-shortcomings | argues-for | inter-task-class-separation-limits-class-IL | class-IL; feature-selection problem — globally discriminative features look locally non-discriminative within one task, inter-task boundaries are not trainable without past data; t-SNE shows only rehearsal disentangles tasks in latent space |
lesort2019-regularization-shortcomings | argues-for | replay-is-necessary-for-class-IL | single-head MNIST-Fellowship; vanilla rehearsal at 100 samples/class succeeds where all tested regularization fails, and the conclusion recommends replay-regularization coupling as the alternative to test-time task labels |
kim2022-theoretical-cil | argues-for | cil-decomposes-into-wp-and-task-ood | theory over any trained CIL model under cross-entropy: H_CIL ≤ ε + δ given WP ≤ ε and TP ≤ δ (Thm 1); TP and per-task OOD detection bound each other up to a constant (Thm 2); good WP and good TP/OOD implicitly exist in any good CIL model (Thm 4) — necessity and sufficiency |
kim2022-theoretical-cil | argues-for | inter-task-class-separation-limits-class-IL | names and defines ICS as CIL's additional difficulty beyond CF — new/old class boundaries cannot be established without old-task data; perfect per-task OOD detection dissolves it, reducing CIL to WP |
kim2022-theoretical-cil | argues-for | task-identity-can-be-inferred-from-inputs | class-IL from scratch, six splits of MNIST/CIFAR/TinyImageNet: task-id read off per-task OOD scores with no task oracle and no stored exemplars; HAT+CSI 63.3 vs best replay baseline DER++ 53.7 on C100-10T, 87.8 vs 66.0 on C10-5T |
kim2022-theoretical-cil | nuances | task-identity-can-be-inferred-from-inputs | what governs the trick is OOD-detection strength, not the inference heuristic: same backbone jumps 41.2→63.3 CIL on C100-10T when ODIN is swapped for CSI, and prior task-id-inference systems — Expert Gate 43.2 on M-5T, iTAML 33.5 on C100-10T under single-instance testing — are weak precisely for lacking it |
kim2022-theoretical-cil | nuances | scenario-choice-determines-conclusions | theory: the TIL–CIL gap is given exact content — TIL = WP and CIL = WP × TP — so the scenario difference is precisely the task-id/OOD term rather than an unstructured empirical gap, and TIL methods become CIL methods by adding per-task OOD detection |
kim2022-theoretical-cil | argues-against | replay-is-necessary-for-class-IL | class-IL from scratch: exemplar-free HAT+CSI/Sup+CSI outperform all replay baselines — 87.8/86.0 vs DER++ 66.0 on C10-5T, 63.3/65.1 vs 53.7 on C100-10T — with the buffer used only for the optional +c output calibration |
kim2022-theoretical-cil | nuances | replay-is-necessary-for-class-IL | reinterpretation, implication section: replay's benefit is read as old-task samples naturally acting as OOD negatives that improve TP — sub-optimally, since the buffer is small and earlier tasks never see future-task data |
kim2022-theoretical-cil | argues-for | regularization-alone-fails-in-class-IL | mechanism reading plus numbers: exemplar-free regularization methods leave TP weak because per-task training never considers OOD, so ICS stays unsolved — MUC 30.4, PASS 33.0 on C100-10T vs 63.3 for HAT+CSI |
kim2022-theoretical-cil | argues-for | dynamic-architectures-sidestep-forgetting | task-IL premise of the recipe: parameter-isolation methods HAT/SupSup described as having almost completely overcome forgetting, which is what reduces CIL to the OOD term; HAT+CSI reaches 92.0 TIL on C100-10T |
kim2022-theoretical-cil | argues-for | low-forgetting-can-mask-low-plasticity | appendix forgetting-rate analysis: OWM and HyperNet show lower forgetting rates than HAT+CSI+c on split-TinyImageNet yet reach only 10.0/7.9 accuracy on T-5T vs 51.7 — described as unable to adapt to new classes |
kim2022-theoretical-cil | builds-on | kim2022-clom | |
kim2022-theoretical-cil | builds-on | serra2018-hat | |
kim2022-theoretical-cil | builds-on | wortsman2020-supsup | |
kim2022-theoretical-cil | builds-on | tack2020-csi | |
kim2022-theoretical-cil | builds-on | vandeven2019-three-scenarios | |
kim2022-theoretical-cil | argues-for | replay-acts-as-ood-negatives | implication section, no dedicated experiment: replay data "naturally regarded as OOD data for the current task", building a better OOD model and hence better TP; sub-optimal because the buffer is small and earlier tasks never see future-task data |
lin2022-trgp | builds-on | saha2021-gpm | |
lin2022-trgp | argues-for | orthogonal-updates-preserve-past-tasks | task-IL, from scratch; strict orthogonal projection against non-trust-region old-task input subspaces still carries retention — BWT −0.8/−0.9/−0.04/−0.5 on PMNIST/CIFAR-100-Split/5-Datasets/MiniImageNet |
lin2022-trgp | first-mention | trainable-subspace-needs-new-task-alignment | task-IL, from scratch; two toy cases show orthogonal projection zeroes the update when new-task and old-task subspaces are parallel — the trainable directions must overlap the new task's own subspace, motivating correlation-based reuse; TRGP 96.34/74.46/93.56/61.78 ACC vs GPM 93.91/72.48/91.22/60.41 |
lin2022-trgp | nuances | capacity-equals-projector-rank | task-IL, from scratch; GPM's remaining capacity is exhausted by projector rank as old subspaces accumulate — TRGP recovers plasticity on late/"difficult" tasks not by adding rank but by reusing correlated old subspaces through learned scaling matrices, largest gains on the hard tail tasks e.g. CIFAR-10/SVHN/notMNIST wi... |
lin2022-trgp | argues-for | frozen-module-access-enables-forward-transfer | task-IL, from scratch; reusing frozen old-task weight projections via a learned scaling matrix Qⱼ,ₜˡ improves forward transfer over strict freezing — +2.43/+1.98/+1.37% ACC over GPM on PMNIST/CIFAR-100-Split/MiniImageNet |
lin2022-trgp | tests | forgetting-is-non-monotonic-in-task-similarity | task-IL, from scratch; TRGP conditions reuse on measured per-layer input-subspace correlation between the new and old tasks — strongly correlated tasks reused, weakly correlated ones protected orthogonally |
cermelli2020-mib | first-mention | background-class-semantics-shift-across-ciss-steps | CISS, Pascal-VOC 2012 & ADE20K, DeepLab-v3/ResNet-101; each step labels only current classes so past/future classes collapse into a background whose meaning shifts step to step, aggravating forgetting |
cermelli2020-mib | argues-for | background-class-semantics-shift-across-ciss-steps | VOC, ADE20K; modeling the shift in the CE/KD losses and classifier init yields ~15% mIoU over LwF on VOC and +11.6% over the best baseline LwF-MC on ADE20K 100-10 all-classes |
cermelli2020-mib | argues-for | knowledge-distillation-preserves-past-knowledge | CISS, VOC & ADE20K; data-focused distillation baselines LwF/LwF-MC/ILT beat prior-focused EWC/PI/RW by a large margin, and MiB's background-aware distillation extends this to segmentation |
cermelli2020-mib | nuances | knowledge-distillation-preserves-past-knowledge | CISS with a shared background class; a naive distillation port is insufficient because the old and current models' background probabilities carry different semantics — distillation must be redefined over the shifting label space, e.g. MiB 15-1 VOC overlapped all-mIoU 29.7 vs ILT 5.7 |
cermelli2020-mib | argues-for | scenario-choice-determines-conclusions | CISS, VOC & ADE20K; the overlapped protocol is consistently the hardest and rankings depend on the disjoint-vs-overlapped image-sampling choice |
cermelli2020-mib | builds-on | michieli2019-ilt | |
cermelli2020-mib | builds-on | li2016-lwf | |
huszar2018-quadratic-penalties | builds-on | kirkpatrick2017-ewc | |
huszar2018-quadratic-penalties | argues-for | posterior-becomes-prior-retains-knowledge | analytical: recursive Bayes / diagonalized Laplace propagation over a task sequence; EWC recovered as its point-estimate, two-task special case |
huszar2018-quadratic-penalties | nuances | protecting-important-weights-prevents-forgetting | analytical: the importance penalty should be a single running penalty anchored at the latest solution with a moving-sum Fisher, not one penalty per past task — EWC's per-task sum double-counts earlier data and biases toward earlier tasks; constant storage in #tasks |
huszar2018-quadratic-penalties | first-mention | single-running-penalty-avoids-double-counting | |
huszar2018-quadratic-penalties | argues-for | single-running-penalty-avoids-double-counting | analytical, recursive Laplace over a task sequence: EWC's per-task penalty sum is consistent only for two tasks and double-counts earlier data from the third on; the consistent form is one running penalty at the latest solution with a moving-sum Fisher, at storage constant in #tasks |
cudrano2024-continual-visual-odometry | argues-for | static-benchmark-results-dont-predict-embodied-deployment | embodied VO, Habitat, scene-level, from scratch; standard-benchmark CL methods EWC/LwF/replay fail to mitigate forgetting, naive final loss 3.88 vs joint 0.50 ×10⁻³ |
cudrano2024-continual-visual-odometry | argues-against | protecting-important-weights-prevents-forgetting | embodied VO regression, Habitat, from scratch; EWC and LwF track naive fine-tuning — final loss 4.29 / 4.03 vs naive 3.88 ×10⁻³, no mitigation |
cudrano2024-continual-visual-odometry | nuances | replay-is-necessary-for-class-IL | embodied VO regression, domain-IL not class-IL; random replay only partially closes the gap even at a full-apartment buffer — final 2.26 vs naive 3.88, joint 0.50 ×10⁻³ |
cudrano2024-continual-visual-odometry | nuances | replay-selection-matters-at-small-budgets | embodied VO, Habitat; uniform random replay's benefit scales with buffer size 1024/5120/13888 → 3.39/2.69/2.26 ×10⁻³, no selection heuristic tested but plain random gains track budget |
cudrano2024-continual-visual-odometry | tests | evaluation-should-go-beyond-forgetting | embodied VO; measures forward transfer and specialization jointly with forgetting over the 72-apartment stream |
cudrano2024-continual-visual-odometry | first-mention | embodied-cl-specializes-after-early-transfer | continual visual odometry, agent stream; an early-transfer-then-specialize pattern where later forgetting is active over-fitting to the present rather than passive decay |
cudrano2024-continual-visual-odometry | first-mention | action-conditioning-trades-convergence-for-specialization | continual VO ablation; conditioning on the agent's actions speeds convergence on the current stream but increases specialization and forgetting |
rosseau2026-dynamical-isometry | argues-for | networks-lose-plasticity-under-continual-training | from scratch, 200–1000-task supervised streams — permuted MNIST/CIFAR-10, random-label CIFAR-10, label-shuffled CIFAR-100 — and continual RL, MinAtar/Octax with PPO, where the unmodified baseline progressively loses the ability to learn |
rosseau2026-dynamical-isometry | nuances | networks-lose-plasticity-under-continual-training | reinterprets the idea's separate correlates — dead units, weight-norm growth, rank collapse — as facets of a single quantity: layer-wise Jacobian singular values drifting from one, i.e. loss of dynamical isometry / empirical-NTK anisotropy |
rosseau2026-dynamical-isometry | builds-on | dohare2024-plasticity | |
rosseau2026-dynamical-isometry | builds-on | lewandowski2025-spectral | |
rosseau2026-dynamical-isometry | argues-for | dynamical-isometry-preserves-plasticity | from scratch: permuted MNIST/CIFAR-10, random-label CIFAR-10, label-shuffled CIFAR-100, 200–1000 tasks, MLPs/CNNs/small transformer; continual MinAtar and Octax with PPO; NTK-conditioning, Jacobian-spectrum and dormant-unit diagnostics |
rosseau2026-dynamical-isometry | nuances | spectral-norm-control-suffices-for-trainability | reads σ_max-only regularization as a partial slice of the isometry target — the Gram-deviation penalty's largest term; argues the full spectrum, i.e. NTK conditioning, is the operative quantity |
rosseau2026-dynamical-isometry | nuances | dormant-neuron-accumulation-drives-plasticity-loss | uses dormant-unit counts as a diagnostic but argues ReDo-style resets can exacerbate NTK ill-conditioning |
hu2025-continual-knowledge-rl | argues-for | task-vector-merging-enables-rehearsal-free-cl | continual RL, from scratch, Meta-World/Freeway/SpaceInvaders; per-task knowledge vectors relative to a frozen base, no stored transitions; +8.02% forward transfer over CompoNet |
hu2025-continual-knowledge-rl | nuances | task-vector-merging-enables-rehearsal-free-cl | merging is folded into sequential training with per-task *learned* softmax composition weights + a fresh task vector — not a one-shot merge of independently-trained checkpoints replacing sequential training; naive averaging without learned adaptation hurts, 0.7532 → 0.7437 |
hu2025-continual-knowledge-rl | argues-for | evaluation-should-go-beyond-forgetting | continual RL; adaptive merging holds parameter memory and inference latency ~constant past ~5 tasks vs linear growth for CompoNet/ProgNet, and forward transfer is the headline metric — forgetting kept balanced, not minimized |
vahedifar2026-nfl | argues-for | knowledge-distillation-preserves-past-knowledge | task-IL and class-IL, split-CIFAR-100/split-TinyImageNet/split-ImageNet, ResNet-18 from scratch; retention carried by dual soft-target anchoring — distilling against both old logits and intermediate old-head-on-adapted-features logits — CIFAR-100 10-task class-IL 53.70% ACC at BWT −0.05 |
vahedifar2026-nfl | argues-for | consolidation-phases-decouple-transfer-from-retention | same setting; a fixed 4-stage schedule alternates an unconstrained new-head learning phase with a distillation-based backbone-consolidation phase over one fixed network — no separate protected store or growth |
vahedifar2026-nfl | argues-for | dynamic-architectures-sidestep-forgetting | fixed-capacity variant without growth or masks: isolation comes from the stepwise freeze schedule; ResNet-18 from scratch, requires boundaries and train-time task identity for per-task heads |
vahedifar2026-nfl | argues-against | replay-is-necessary-for-class-IL | class-IL, split-CIFAR-100/split-TinyImageNet/split-ImageNet, ResNet-18 from scratch: buffer-free NFL+ leads all buffer-free baselines and matches memory-based methods at 2.53% of their model size — CIFAR-100 10-task 53.70% vs DyTox 55.80% / MEMO 54.40%; trails DyTox by ~1.7–2.1pp on ImageNet-1000; task identity still u... |
vahedifar2026-nfl | argues-for | regularization-alone-fails-in-class-IL | class-IL, split-ImageNet 10–50 tasks, ResNet-18 from scratch: EWC/SI/LwF fall far behind buffer-free NFL+ and the memory-based methods |
vahedifar2026-nfl | builds-on | kang2022-wsn | |
elsayed2024-upgd | first-mention | jointly-addressing-plasticity-and-forgetting-via-unit-utility | streaming, unknown task boundaries, no replay; UPGD gates each weight's gradient and perturbation by a utility = counterfactual loss-change; among the few methods to keep improving on both label-permuted EMNIST and mini-ImageNet |
elsayed2024-upgd | argues-for | networks-lose-plasticity-under-continual-training | streaming input-permuted MNIST, 200 tasks; SGD/Adam accuracy decays over tasks and falls below Adam-Restarts, distinct from forgetting |
elsayed2024-upgd | argues-for | protecting-important-weights-prevents-forgetting | streaming, boundary-free; protecting high-utility weights from gradient overwrite is what supplies retention — the non-protecting ablation forgets |
elsayed2024-upgd | tests | dormant-neuron-accumulation-drives-plasticity-loss | input-permuted MNIST; % zero activations rises and gradient ℓ0-norm falls for all methods except UPGD-W, non-protecting UPGD-W, and S&P |
elsayed2024-upgd | builds-on | dohare2024-plasticity | |
zheng2023-learn-or-recall | argues-for | forgetting-is-concentrated-in-higher-layers | class-IL + task-IL NLP classification, PLMs 19M–1.21B, encoder- and decoder-only: backbone linear-probing accuracy barely drops after task 1 while observed accuracy collapses 98%→10% on split-CLINC-150; the forgetting is located in the classifier, whose old class embeddings drift from their learned positions |
zheng2023-learn-or-recall | argues-for | task-recency-bias-dominates-class-il-errors | class-IL, PLM backbones under SEQ: new-class logits dominate old ones and the failure is correctable entirely at the classifier level — freezing old classifiers + cosine head lifts split-CLINC-150 from 9.3 to 64.3 without touching the backbone |
zheng2023-learn-or-recall | nuances | task-recency-bias-dominates-class-il-errors | Pythia-410m, split-CLINC-150: the bias is not a norm artifact — new tasks' class-embedding norms are no larger than old tasks' — but a geometric drift of old class embeddings away from their optimal position relative to class feature centres |
zheng2023-learn-or-recall | argues-for | simple-baselines-rival-sophisticated-methods | class-IL/task-IL NLP without stored samples: SEQ* — warm-up, freeze backbone and old classifiers, cosine/linear head — matches or beats 20+ SOTA IL methods, e.g. 84.5 vs best SOTA 47.1 on split-CLINC-150, with 10.24K trainable params per task vs 405M and 24 vs 68–199 min training |
zheng2023-learn-or-recall | argues-for | first-session-adaptation-suffices | PLM class-IL: full-parameter warm-up for 1–3 epochs on the first task only, then freezing, brings considerable improvement when the pretrain–downstream gap is large — split-CLINC-150, split-Banking77, split-FewRel, split-TACRED — and little when it is small — Topic3Datasets |
zheng2023-learn-or-recall | argues-for | masked-output-updates-protect-past-classes | hard variant — freezing old-task classifier weights outright rather than masking their losses: SEQ FixC+Cos reaches 64.3 vs SEQ Lin 9.3 on split-CLINC-150, and "freezing old classifiers is more important than freezing backbones" |
zheng2023-learn-or-recall | argues-for | future-head-preparation-aids-class-il | generative PLM backbones, class-IL: pre-allocating all future classifiers from the start lifts SEQ* from 76.0 to 84.5 on split-CLINC-150, attributed to reduced overlap between class embeddings; marked optional because it requires knowing the total task count |
zheng2023-learn-or-recall | argues-for | frozen-features-plus-cheap-classifier-rivals-full-cl | NLP class-IL with PLM backbones: a backbone frozen after first-task warm-up plus a 10.24K-parameter gradient-trained linear/cosine head — not closed-form — rivals or beats 20+ full CL methods at a fraction of time and parameters |
zheng2023-learn-or-recall | argues-for | pretraining-mitigates-forgetting | probing across Pythia pretraining checkpoints: pretraining raises the backbone's linear-probing accuracy in IL, with the largest gains on harder tasks such as relation extraction; larger checkpoints adapt to downstream tasks without the pretrain–downstream gap smaller ones show |
zheng2023-learn-or-recall | nuances | pretraining-mitigates-forgetting | the backbone's anti-forgetting is not from pretraining alone: randomly initialized transformers reach high linear-probing accuracy under SEQ on intent classification, and SEQ improves probing accuracy from nearly all pretraining checkpoints — the Transformer architecture is a co-factor |
zheng2023-learn-or-recall | argues-for | pretraining-overlap-invalidates-cil-benchmarks | SEQ* wins where the PLM plausibly already holds the knowledge and loses its edge where new knowledge must be absorbed — fine-grained split-Few-NERD entities — or classes overlap across tasks — Topic3Datasets; the conclusion calls for IL benchmarks requiring domain-specific knowledge |
zheng2023-learn-or-recall | nuances | solver-can-be-its-own-replay-generator | PLM class-IL NLP: LAMOL-family self-generated rehearsal works when downstream data is close to the pretraining distribution — LAMOL_t 74.1 vs SEQ* 70.6 on Topic3Datasets — but degrades on longer sequences and larger gaps — 43.4 vs 84.5 on split-CLINC-150, 28.4 vs 62.0 on split-FewRel; LFPT5's soft-prompt-only generator... |
zheng2023-learn-or-recall | builds-on | davari2022-probing | |
zheng2023-learn-or-recall | builds-on | tao2023-bert-probing | |
zheng2023-learn-or-recall | first-mention | plm-anti-forgetting-is-partly-architectural | linear probing of Pythia checkpoints at pretraining steps 0–143k, class-IL intent classification / relation extraction: random-init backbones reach high probing accuracy on intent classification, SEQ improves probing from nearly all checkpoints; pretraining still needed for harder tasks |
zheng2023-learn-or-recall | first-mention | probe-choice-governs-measured-retention | four probing metrics across 8 NLP datasets, class-IL/task-IL, BERT/GPT2/Pythia: linear > cosine-linear ≈ cosine-prototype > prototype, explained by feature-cone anisotropy and norm-borne pretraining knowledge |
chaudhry2019-agem | builds-on | lopezpaz2017-gem | |
chaudhry2019-agem | argues-for | evaluation-should-go-beyond-forgetting | single-pass 20-task streams from scratch: learning speed — the LCA_10 metric — plus training time and memory reported as first-class axes beside accuracy/forgetting; A-GEM matches GEM's accuracy, 62.3 vs 61.2 on split-CIFAR-100, at ~100× less time, 449s vs 5238s, and ~10× less memory, 2P vs P·T growth — a difference in... |
chaudhry2019-agem | nuances | hyperparameters-must-be-set-from-current-task-only | shares the premise — whole-stream grid search "strictly speaking violat[es] the assumption of the LLL scenario" — but prescribes a different compliant remedy: multi-pass cross-validation on the first T_CV=3 tasks, disjoint from the 17 single-pass evaluation tasks; under it PI fails to train (∞) on permuted-MNIST |
chaudhry2019-agem | nuances | protecting-important-weights-prevents-forgetting | single-pass from-scratch 20-task streams: EWC 42.4 vs naive 42.9 on split-CIFAR-100; Appendix F epoch/width sweeps locate EWC's gains in the multi-epoch, over-parameterized regime — importance estimated from an unconverged model protects little |
chaudhry2019-agem | nuances | inequality-loss-constraints-permit-positive-backward-transfer | the inequality constraints need not be per-task: a single constraint on the average memory loss retains average accuracy — 62.3 vs GEM's 61.2 on split-CIFAR-100 — while giving up GEM's per-task worst-case guarantee, worst-case forgetting 0.15 vs 0.05 on the stored CIFAR memories |
chaudhry2019-agem | first-mention | averaged-gradient-constraint-matches-per-task-constraints | |
chaudhry2019-agem | argues-for | averaged-gradient-constraint-matches-per-task-constraints | task-IL online single-pass 20-task streams from scratch: A-GEM 62.3 vs GEM 61.2 on split-CIFAR-100, 89.1 vs 89.5 on permuted-MNIST; training time 449s vs 5238s; memory 2P vs P·T |
chaudhry2019-agem | nuances | averaged-gradient-constraint-matches-per-task-constraints | own Appendix D.2: worst-case per-task forgetting is surrendered — Fwst 0.15 vs GEM 0.05, accuracy on stored CIFAR memories 72.1 vs 97.1 |
chaudhry2019-agem | first-mention | compositional-task-descriptors-enable-zero-shot-transfer | |
chaudhry2019-agem | argues-for | compositional-task-descriptors-enable-zero-shot-transfer | task-IL online, split-CUB/split-AWA, attribute-matrix descriptors: joint-embedding variants lift every method — VAN 54.3→67.1, EWC 54→68.4, A-GEM 62→71 A_T on CUB — and zero-shot accuracy rises over the learning experience |
End of preview. Expand in Data Studio
The Continual-Learning Ideas Vault
A discourse graph of the continual-learning literature, built with LLM agents. Papers and ideas are notes, linked by typed stance edges (first-mention, argues-for, argues-against, nuances, tests, builds-on) that carry the experimental conditions inline. One edge, as stored:
french1999-catastrophic-forgetting—argues-for→generative-replay-can-substitute-stored-data("pseudo-rehearsal / pseudopatterns, feedforward BP nets, no stored past data; Robins 1995 + French 1997 pseudo-recurrent dual-network transfer")
A paper takes a stance on an idea, and the conditions of the claim travel
with the link. This snapshot holds 549 papers, 683 ideas and 3,913 edges;
each edge type is defined in PROJECT.md.
Stance: descriptive, not evaluative. The map records what papers claim and debate, never whether they are right. It has not been systematically checked: errors remain, so verify important claims against the papers.
PDFs are not included. Each note gives the paper's arXiv ID, and
python3 _scripts/fetch_pdfs.py downloads them into pdf/. Until then,
_scripts/check_integrity.py reports each paper as pdf-missing.
Structure
papers/: one note per analyzed paper (metadata + typed edges)ideas/: one falsifiable claim per note, linked to its parent ideamethods/,concepts/,syntheses/,influences/,authors/: methods, core concepts, per-idea landscapes, out-of-scope influences, author pages_ideas-index.md,_methods-index.md,references.bib: generated indexesPROJECT.md: the spec (scope, stance, conventions);CL.md: what continual learning means here;_vocab.md: controlled vocabularyskills/,_scripts/: the agent procedures and scripts that build and check the vaulttables/: the graph as JSONL (notes.jsonl,edges.jsonl), for the dataset viewer, DuckDB or pandasLICENSES/: full license texts
Use it
- Get it:
git clone https://huggingface.co/datasets/imecai/cl-agentic-vault(or pick files from the Files tab above). - Open the folder as an Obsidian vault, grep it, or point your agent at it; the skills show how it was built and checked.
- Query the tables, e.g. the ideas most argued against:
duckdb -c "SELECT target, COUNT(*) FROM 'tables/edges.jsonl' WHERE kind='argues-against' GROUP BY 1 ORDER BY 2 DESC" - Contribute: corrections and additions are welcome as issues or pull requests.
License
Copyright (c) 2026 imec.
- Notes, tables and figures: CC BY 4.0 (
LICENSES/CC-BY-4.0.txt). - Code (
_scripts/,skills/): MIT (LICENSES/MIT.txt).
The papers remain under their own licenses and are not included; short quoted passages and paper titles belong to their authors.
- Downloads last month
- 779